Models
Open weights
An open-weights model has downloadable parameters, so you can run it on your own hardware instead of calling somebody’s API.
Also written: open source model, open weight model
The appeal is control: no per-token pricing, no rate limits, no data leaving your infrastructure, and no model deprecation on somebody else’s timeline.
The cost moves rather than disappearing. You now pay for GPUs, which bill by the hour whether or not you are using them, plus the engineering to run inference reliably.
In practice
Open weights change the cost structure, not the cost to zero: you stop paying per token and start paying for GPUs that are idle between requests. The crossover depends almost entirely on utilisation. Below steady, heavy traffic an API is usually cheaper; the other reasons — data residency, no third-party disclosure, no deprecation — often matter more than the arithmetic.
Common questions
Is self-hosting cheaper than an API?
Only at steady, heavy utilisation. You stop paying per token and start paying for GPUs that are idle between requests, so the crossover depends almost entirely on how full those GPUs stay.
What are the non-cost reasons to use open weights?
Data never leaving your infrastructure, no third-party disclosure to document, no deprecation timetable set by someone else, and the ability to fine-tune without a provider's permission or format.